On Tue, 11 Oct 2022 17:30:02 GMT, Xue-Lei Andrew Fan <xue...@openjdk.org> wrote:
>> This patch rewrites some BigInteger and curve point operations used in EC
>> calculations:
>> - coefficient * 2^power is equivalent to coefficient << power
>> - number mod 2^n is equivalent to number & (2^n-1)
>> - pair of IntegerModuloP operations:
>> t2 = t1+t1
>> t1 = t1+t2
>> is equivalent to t1=t1*3, which is now implemented more efficiently.
>>
>> Benchmarked the code using not-yet-merged benchmark from #10544. Results on
>> x64 before:
>>
>> Benchmark (messageLength) Mode Cnt Score Error Units
>> Signatures.sign 64 thrpt 15 1578.907 ± 1.522 ops/s
>>
>> After:
>>
>> Benchmark (messageLength) Mode Cnt Score Error Units
>> Signatures.sign 64 thrpt 15 1679.495 ± 3.883 ops/s
>>
>> Greatest part of the improvement is related to ECOperations changes;
>> BigInteger modifications provide only marginal gains (1584 ops/s without
>> ECOperations changes).
>>
>> Tier1-3 tests continue to pass.
>
> src/java.base/share/classes/sun/security/provider/DSAParameterGenerator.java
> line 240:
>
>> 238: W = W.add(V[i].shiftLeft(i * outLen));
>> 239: }
>> 240: W = W.add((V[n].mod(BigInteger.TWO.pow(b)))
>
> Did you want to update the "BigInteger.TWO.pow(b)" as well?
Actually I think I'll revert the DSA changes; I'm not prepared to measure their
effect at this moment.
> src/java.base/share/classes/sun/security/util/math/intpoly/IntegerPolynomial.java
> line 332:
>
>> 330:
>> 331: protected void setLimbsValuePositive(BigInteger v, long[] limbs) {
>> 332: assert bitsPerLimb < 32;
>
> I may have this assert in the constructors as it is a final field.
I added it here as a comment explaining why using intValue is okay here; it
wouldn't serve its purpose if I moved it elsewhere.
Asserts are eliminated by the compiler by default anyway. If you think this
assert shouldn't be here, I can replace it with a normal comment.
-------------
PR: https://git.openjdk.org/jdk/pull/10614
